Image Understanding on TextVQA (val)
88.39AccuracyVanilla MoE
Evaluation Results
| Method | Links | |
|---|---|---|
| Vanilla MoEBackbone=KIMI-VL-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 88.39 | |
| MACS (Ours)Backbone=KIMI-VL-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 88.27 | |
| MACS (w/o Expanded)Backbone=KIMI-VL-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 87.94 | |
| Vanilla MoEBackbone=INTERNVL3.5-30B-A3B, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 85.68 | |
| MACS (Ours)Backbone=INTERNVL3.5-30B-A3B, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 85.51 | |
| MACS (w/o Expanded)Backbone=INTERNVL3.5-30B-A3B, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 84.92 | |
| CAI-MoE (Expanded)Backbone=KIMI-VL-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 84.72 | |
| Vanilla MoEBackbone=QWEN3-VL-30B-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 83.54 | |
| MACS (Ours)Backbone=QWEN3-VL-30B-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 83.41 | |
| MACS (w/o Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 83.04 | |
| CAI-MoE (Token Drop)Backbone=KIMI-VL-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 82.56 | |
| CAI-MoE (Expanded)Backbone=INTERNVL3.5-30B-A3B, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 80.94 | |
| CAI-MoE (Expanded)Backbone=QWEN3-VL-30B-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 79.86 | |
| CAI-MoE (Token Drop)Backbone=INTERNVL3.5-30B-A3B, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 78.13 | |
| CAI-MoE (Token Drop)Backbone=QWEN3-VL-30B-A3B-INSTRUCT, Base Capacity Factor=1.0, Evaluation Protocol=EP inference setting2026.04 | 77.42 |